LLM inference in C/C++
C++ ggml
model: support embeddinggemma2 (text+vision+audio) (#30054)
Xuan-Son Nguyen4fbc76dec51d
diff --git a/conversion/__init__.py b/conversion/__init__.pyindex 7ba79a3c50..d8af76a270 100644--- a/conversion/__init__.py+++ b/conversion/__init__.py@@ -74,6 +74,7 @@ TEXT_MODEL_MAP: dict[str, str] = {"Dots3NoteTextForCausalLM": "dots3","DotsOCRForCausalLM": "qwen","DreamModel": "dream",+ "EmbeddingGemma2Model": "gemma","Ernie4_5ForCausalLM": "ernie","Ernie4_5_ForCausalLM": "ernie","Ernie4_5_MoeForCausalLM": "ernie",@@ -308,6 +309,7 @@ MMPROJ_MODEL_MAP: dict[str, str] = {"Dots3NoteForCausalLM": "dots3","Dots3NoteForConditionalGeneration": "dots3","DotsOCRForCausalLM": "dotsocr",+ "EmbeddingGemma2Model": "gemma","Exaone4_5_ForConditionalGeneration": "exaone","Gemma3ForConditionalGeneration": "gemma","Gemma3nForConditionalGeneration": "gemma",diff --git a/conversion/gemma.py b/conversion/gemma.pyindex 9ec622ed49..2a1f6931ff 100644--- a/conversion/gemma.py+++ b/conversion/gemma.py@@ -700,7 +700,7 @@ class Gemma4Model(Gemma3Model):self.gguf_writer.add_key_length_swa(head_dim_swa)self.gguf_writer.add_value_length_swa(head_dim_swa)- expert_intermediate_size = self.find_hparam(["expert_intermediate_size", "moe_intermediate_size"])+ expert_intermediate_size = self.find_hparam(["expert_intermediate_size", "moe_intermediate_size"], optional=True)if expert_intermediate_size is not None:self.gguf_writer.add_expert_feed_forward_length(expert_intermediate_size)@@ -810,6 +810,28 @@ class Gemma4Model(Gemma3Model):yield from super().modify_tensors(data_torch, name, bid)[email protected]("EmbeddingGemma2Model")+# TODO: add example model+class EmbeddingGemma2Model(Gemma4Model):+ model_arch = gguf.MODEL_ARCH.GEMMA_EMBEDDING2++ def __init__(self, *args, **kwargs):+ super().__init__(*args, **kwargs)+ self.hparams["num_kv_shared_layers"] = 0++ def set_gguf_parameters(self):+ super().set_gguf_parameters()+ # HF sliding_window is bidirectional, llama.cpp expects the full window size+ self.gguf_writer.add_sliding_window(2 * self.hparams["sliding_window"])+ self.gguf_writer.add_embedding_length_out(self.hparams["embedding_dim"])+ self.gguf_writer.add_causal_attention(False)+ self._try_set_pooling_type()++ def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:+ # default rope on all layers, no rope_freqs needed+ return iter(())++@ModelBase.register("Gemma4DSparkModel")class Gemma4DSparkModel(DFlashModel):model_arch = gguf.MODEL_ARCH.DFLASH@@ -1030,6 +1052,14 @@ class Gemma4VisionAudioModel(MmprojModel):yield (mapped_name, data_torch)[email protected]("EmbeddingGemma2Model")+# TODO: add example model+class EmbeddingGemma2VisionAudioModel(Gemma4VisionAudioModel):+ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:+ # same towers as Gemma4, but the tensor names have no "model." prefix+ yield from super().modify_tensors(data_torch, "model." + name, bid)++@ModelBase.register("Gemma4UnifiedForConditionalGeneration")@ModelBase.example("hf-tiny-v2/tiny-random-Gemma4UnifiedForConditionalGeneration")class Gemma4UnifiedVisionAudioModel(Gemma4VisionAudioModel):diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.pyindex 6b872da06d..ea7ddfa576 100644--- a/gguf-py/gguf/constants.py+++ b/gguf-py/gguf/constants.py@@ -559,6 +559,7 @@ class MODEL_ARCH(IntEnum):GEMMA4 = auto()GEMMA4_ASSISTANT = auto()GEMMA_EMBEDDING = auto()+ GEMMA_EMBEDDING2 = auto()STARCODER2 = auto()RWKV6 = auto()RWKV6QWEN2 = auto()@@ -1338,6 +1339,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {MODEL_ARCH.GEMMA4: "gemma4",MODEL_ARCH.GEMMA4_ASSISTANT: "gemma4-assistant",MODEL_ARCH.GEMMA_EMBEDDING: "gemma-embedding",+ MODEL_ARCH.GEMMA_EMBEDDING2: "gemma-embedding2",MODEL_ARCH.STARCODER2: "starcoder2",MODEL_ARCH.RWKV6: "rwkv6",MODEL_ARCH.RWKV6QWEN2: "rwkv6qwen2",@@ -3471,6 +3473,30 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {MODEL_TENSOR.FFN_PRE_NORM,MODEL_TENSOR.FFN_POST_NORM,],+ MODEL_ARCH.GEMMA_EMBEDDING2: [+ MODEL_TENSOR.TOKEN_EMBD,+ MODEL_TENSOR.OUTPUT,+ MODEL_TENSOR.OUTPUT_NORM,+ MODEL_TENSOR.ATTN_Q,+ MODEL_TENSOR.ATTN_Q_NORM,+ MODEL_TENSOR.ATTN_K,+ MODEL_TENSOR.ATTN_K_NORM,+ MODEL_TENSOR.ATTN_V,+ MODEL_TENSOR.ATTN_OUT,+ MODEL_TENSOR.FFN_GATE,+ MODEL_TENSOR.FFN_DOWN,+ MODEL_TENSOR.FFN_UP,+ MODEL_TENSOR.ATTN_NORM,+ MODEL_TENSOR.ATTN_POST_NORM,+ MODEL_TENSOR.FFN_PRE_NORM,+ MODEL_TENSOR.FFN_POST_NORM,+ MODEL_TENSOR.LAYER_OUT_SCALE,+ MODEL_TENSOR.PER_LAYER_MODEL_PROJ,+ MODEL_TENSOR.PER_LAYER_INP_GATE,+ MODEL_TENSOR.PER_LAYER_PROJ,+ MODEL_TENSOR.PER_LAYER_PROJ_NORM,+ MODEL_TENSOR.PER_LAYER_POST_NORM,+ ],MODEL_ARCH.STARCODER2: [MODEL_TENSOR.TOKEN_EMBD,MODEL_TENSOR.OUTPUT_NORM,diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.pyindex cef04d0827..5ac11cb46c 100644--- a/gguf-py/gguf/tensor_mapping.py+++ b/gguf-py/gguf/tensor_mapping.py@@ -88,6 +88,7 @@ class TensorNameMap:"model.lm_head", # dflash"model.transformer.ff_out", # llada"head.decoder", # modern-bert+ "embedding_projection", # embeddinggemma2),MODEL_TENSOR.DENSE_2_OUT: ("dense_2_out", # embeddinggemma@@ -753,6 +754,7 @@ class TensorNameMap:MODEL_TENSOR.LAYER_OUT_SCALE: ("model.layers.{bid}.layer_scalar", # gemma4+ "layers.{bid}.layer_scalar", # embeddinggemma2"model.blocks.{bid}.embed_skip.a_g", # talkie),@@ -762,10 +764,12 @@ class TensorNameMap:MODEL_TENSOR.PER_LAYER_MODEL_PROJ: ("model.per_layer_model_projection", # gemma3n+ "ple.per_layer_model_projection", # embeddinggemma2),MODEL_TENSOR.PER_LAYER_PROJ_NORM: ("model.per_layer_projection_norm", # gemma3n+ "ple.per_layer_projection_norm", # embeddinggemma2),MODEL_TENSOR.ALTUP_PROJ: (@@ -778,14 +782,17 @@ class TensorNameMap:MODEL_TENSOR.PER_LAYER_INP_GATE: ("model.layers.{bid}.per_layer_input_gate", # gemma3n+ "layers.{bid}.ple_block.per_layer_input_gate", # embeddinggemma2),MODEL_TENSOR.PER_LAYER_PROJ: ("model.layers.{bid}.per_layer_projection", # gemma3n+ "layers.{bid}.ple_block.per_layer_projection", # embeddinggemma2),MODEL_TENSOR.PER_LAYER_POST_NORM: ("model.layers.{bid}.post_per_layer_input_norm", # gemma3n+ "layers.{bid}.ple_block.post_per_layer_input_norm", # embeddinggemma2),MODEL_TENSOR.ALTUP_CORRECT_COEF: (diff --git a/src/llama-arch.cpp b/src/llama-arch.cppindex 850737ecaa..e1014ed387 100644--- a/src/llama-arch.cpp+++ b/src/llama-arch.cpp@@ -60,6 +60,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {{ LLM_ARCH_GEMMA4, "gemma4" },{ LLM_ARCH_GEMMA4_ASSISTANT, "gemma4-assistant" },{ LLM_ARCH_GEMMA_EMBEDDING, "gemma-embedding" },+ { LLM_ARCH_GEMMA_EMBEDDING2, "gemma-embedding2" },{ LLM_ARCH_STARCODER2, "starcoder2" },{ LLM_ARCH_MAMBA, "mamba" },{ LLM_ARCH_MAMBA2, "mamba2" },diff --git a/src/llama-arch.h b/src/llama-arch.hindex 24068fb7a4..80344d3280 100644--- a/src/llama-arch.h+++ b/src/llama-arch.h@@ -65,6 +65,7 @@ enum llm_arch {LLM_ARCH_GEMMA4,LLM_ARCH_GEMMA4_ASSISTANT,LLM_ARCH_GEMMA_EMBEDDING,+ LLM_ARCH_GEMMA_EMBEDDING2,LLM_ARCH_STARCODER2,LLM_ARCH_MAMBA,LLM_ARCH_MAMBA2,diff --git a/src/llama-model.cpp b/src/llama-model.cppindex 97e0cee70e..fa379bae88 100644--- a/src/llama-model.cpp+++ b/src/llama-model.cpp@@ -156,6 +156,8 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_paramsreturn new llama_model_gemma4_assistant(params);case LLM_ARCH_GEMMA_EMBEDDING:return new llama_model_gemma_embedding(params);+ case LLM_ARCH_GEMMA_EMBEDDING2:+ return new llama_model_gemma_embedding2(params);case LLM_ARCH_STARCODER2:return new llama_model_starcoder2(params);case LLM_ARCH_MAMBA:@@ -2371,6 +2373,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,case LLM_ARCH_WAVTOKENIZER_DEC:case LLM_ARCH_MODERN_BERT:case LLM_ARCH_GEMMA_EMBEDDING:+ case LLM_ARCH_GEMMA_EMBEDDING2:case LLM_ARCH_DREAM:case LLM_ARCH_LLADA:case LLM_ARCH_LLADA_MOE:@@ -3152,6 +3155,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {case LLM_ARCH_GEMMA4:case LLM_ARCH_GEMMA4_ASSISTANT:case LLM_ARCH_GEMMA_EMBEDDING:+ case LLM_ARCH_GEMMA_EMBEDDING2:case LLM_ARCH_STARCODER2:case LLM_ARCH_OPENELM:case LLM_ARCH_GPTNEOX:diff --git a/src/models/gemma-embedding2.cpp b/src/models/gemma-embedding2.cppnew file mode 100644index 0000000000..4b77664ad3--- /dev/null+++ b/src/models/gemma-embedding2.cpp@@ -0,0 +1,234 @@+#include "models.h"++void llama_model_gemma_embedding2::load_arch_hparams(llama_model_loader & ml) {+ hparams.swa_type = LLAMA_SWA_TYPE_SYMMETRIC;+ load_swa_pattern(ml, 6);++ hparams.causal_attn = false; // embeddings do not use causal attention+ hparams.f_attention_scale = 1.0f; // same as Gemma4, q_norm makes the scaling unnecessary++ ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);+ ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);+ ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);+ ml.get_key(LLM_KV_EMBEDDING_LENGTH_PER_LAYER, hparams.n_embd_per_layer);++ switch (hparams.n_layer()) {+ case 24: type = LLM_TYPE_0_3B; break;+ default: type = LLM_TYPE_UNKNOWN;+ }+}++void llama_model_gemma_embedding2::load_arch_tensors(llama_model_loader &) {+ LLAMA_LOAD_LOCALS;++ const int64_t n_embd_per_layer = hparams.n_embd_per_layer;+ const int64_t n_embd_out = hparams.n_embd_out();++ if (n_embd_head_k != n_embd_head_v) {+ throw std::runtime_error("EmbeddingGemma2 requires n_embd_head_k == n_embd_head_v");+ }+ if (hparams.n_embd_head_k_swa != hparams.n_embd_head_v_swa) {+ throw std::runtime_error("EmbeddingGemma2 requires n_embd_head_k_swa == n_embd_head_v_swa");+ }++ tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);++ per_layer_model_proj = create_tensor(tn(LLM_TENSOR_PER_LAYER_MODEL_PROJ, "weight", 0), {n_embd, n_embd_per_layer * n_layer}, 0);+ per_layer_proj_norm = create_tensor(tn(LLM_TENSOR_PER_LAYER_PROJ_NORM, "weight", 0), {n_embd_per_layer}, 0);++ output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);+ // projects the final hidden state to the embedding dimension+ output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_embd_out}, 0);++ for (int i = 0; i < n_layer; ++i) {+ auto & layer = layers[i];+ const int64_t n_head = hparams.n_head(i);+ const int64_t n_embd_head = hparams.n_embd_head_k(i);+ const int64_t n_embd_k = hparams.n_embd_k_gqa(i);+ const int64_t n_embd_v = hparams.n_embd_v_gqa(i);++ layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);++ layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head * n_head}, 0);+ layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k}, 0);+ layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v}, 0);+ layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head * n_head, n_embd}, 0);++ layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head}, 0);+ layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head}, 0);+ layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);++ layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);+ layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);+ layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0);+ layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);+ layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);++ layer.per_layer_inp_gate = create_tensor(tn(LLM_TENSOR_PER_LAYER_INP_GATE, "weight", i), {n_embd, n_embd_per_layer}, 0);+ layer.per_layer_proj = create_tensor(tn(LLM_TENSOR_PER_LAYER_PROJ, "weight", i), {n_embd_per_layer, n_embd}, 0);+ layer.per_layer_post_norm = create_tensor(tn(LLM_TENSOR_PER_LAYER_POST_NORM, "weight", i), {n_embd}, 0);++ layer.out_scale = create_tensor(tn(LLM_TENSOR_LAYER_OUT_SCALE, "weight", i), {1u}, 0);+ }+}++std::unique_ptr<llm_graph_context> llama_model_gemma_embedding2::build_arch_graph(const llm_graph_params & params) const {+ return std::make_unique<graph>(*this, params);+}++llama_model_gemma_embedding2::graph::graph(const llama_model & model, const llm_graph_params & params) :+ llm_graph_context(params),+ model(model) {+ ggml_tensor * cur;+ ggml_tensor * inpL;++ // important: do not normalize weights for raw embeddings input (i.e. encoded image embeddings)+ inpL = build_inp_embd(model.tok_embd, sqrtf(n_embd));+ cb(inpL, "inp_scaled", -1);++ // inp_pos - contains the positions+ ggml_tensor * inp_pos = build_inp_pos();++ auto * inp_attn = build_attn_inp_no_cache();++ ggml_tensor * inp_out_ids = build_inp_out_ids();++ // inp_per_layer shape: [n_embd_per_layer, n_tokens, n_layer]+ ggml_tensor * inp_per_layer = build_inp_per_layer(inpL);++ for (int il = 0; il < n_layer; ++il) {+ const int64_t n_embd_head = hparams.n_embd_head_k(il);+ const int64_t n_head = hparams.n_head(il);+ const int64_t n_head_kv = hparams.n_head_kv(il);++ const float freq_base_l = model.get_rope_freq_base(cparams, il);+ const float freq_scale_l = model.get_rope_freq_scale(cparams, il);+ const int n_rot_l = hparams.n_rot(il);++ // norm+ cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);+ cb(cur, "attn_norm", il);++ // self-attention+ {+ ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur, model.layers[il].wq_s);+ ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur, model.layers[il].wk_s);+ ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur, model.layers[il].wv_s);+ cb(Qcur, "Qcur", il);+ cb(Kcur, "Kcur", il);+ cb(Vcur, "Vcur", il);++ Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);+ Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);+ Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);++ Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il);+ Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il);+ Vcur = ggml_rms_norm(ctx0, Vcur, hparams.f_norm_rms_eps);+ cb(Qcur, "Qcur_normed", il);+ cb(Kcur, "Kcur_normed", il);+ cb(Vcur, "Vcur_normed", il);++ Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot_l, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,+ ext_factor, attn_factor, beta_fast, beta_slow);+ Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot_l, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,+ ext_factor, attn_factor, beta_fast, beta_slow);+ cb(Qcur, "Qcur_pos", il);+ cb(Kcur, "Kcur_pos", il);++ cur = build_attn(inp_attn,+ model.layers[il].wo, nullptr, model.layers[il].wo_s,+ Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, hparams.f_attention_scale, il);+ }++ if (il == n_layer - 1 && inp_out_ids) {+ cur = ggml_get_rows(ctx0, cur, inp_out_ids);+ inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);+ }++ cur = build_norm(cur, model.layers[il].attn_post_norm, nullptr, LLM_NORM_RMS, il);+ cb(cur, "attn_post_norm", il);++ ggml_tensor * attn_out = ggml_add(ctx0, cur, inpL);+ cb(attn_out, "attn_out", il);++ // feed-forward network+ cur = build_norm(attn_out, model.layers[il].ffn_norm, nullptr, LLM_NORM_RMS, il);+ cb(cur, "ffn_norm", il);++ cur = build_ffn(cur,+ model.layers[il].ffn_up, nullptr, model.layers[il].ffn_up_s,+ model.layers[il].ffn_gate, nullptr, model.layers[il].ffn_gate_s,+ model.layers[il].ffn_down, nullptr, model.layers[il].ffn_down_s,+ nullptr,+ LLM_FFN_GELU, LLM_FFN_PAR, il);+ cb(cur, "ffn_out", il);++ cur = build_norm(cur, model.layers[il].ffn_post_norm, nullptr, LLM_NORM_RMS, -1);+ cb(cur, "ffn_post_norm", il);++ cur = ggml_add(ctx0, cur, attn_out);++ // per-layer embedding+ {+ ggml_tensor * pe_in = cur;+ cb(cur, "pe_in", il);++ cur = build_lora_mm(model.layers[il].per_layer_inp_gate, cur); // [n_embd_per_layer, n_tokens]+ cur = ggml_gelu(ctx0, cur);++ ggml_tensor * inp_this_layer = ggml_view_2d(ctx0, inp_per_layer,+ inp_per_layer->ne[0], inp_per_layer->ne[1], inp_per_layer->nb[1], il * inp_per_layer->nb[2]);++ if (il == n_layer - 1 && inp_out_ids) {+ inp_this_layer = ggml_get_rows(ctx0, inp_this_layer, inp_out_ids);+ }++ cur = ggml_mul(ctx0, cur, inp_this_layer);+ cur = build_lora_mm(model.layers[il].per_layer_proj, cur); // [n_embd, n_tokens]+ cur = build_norm(cur, model.layers[il].per_layer_post_norm, nullptr, LLM_NORM_RMS, il);+ cb(cur, "per_layer_embd_out", il);++ cur = ggml_add(ctx0, pe_in, cur);+ }++ // layer_scalar+ cur = ggml_mul(ctx0, cur, model.layers[il].out_scale);+ cb(cur, "out_scaled", il);++ cur = build_cvec(cur, il);+ cb(cur, "l_out", il);++ // input for next layer+ inpL = cur;+ }++ cur = build_norm(inpL, model.output_norm, nullptr, LLM_NORM_RMS, -1);+ cb(cur, "result_norm", -1);++ // projecting per token is equivalent to projecting after mean pooling+ cur = build_lora_mm(model.output, cur, model.output_s);+ cb(cur, "result_embd", -1);+ res->t_embd = cur;++ ggml_build_forward_expand(gf, cur);+}++// equivalent to EmbeddingGemma2TextPLE in python code+// this model has no per-layer token embeddings, the per-layer inputs come only from the projection+// inpL shape: [n_embd, n_tokens]+// output shape: [n_embd_per_layer, n_tokens, n_layer]+ggml_tensor * llama_model_gemma_embedding2::graph::build_inp_per_layer(ggml_tensor * inpL) {+ const int64_t n_embd_per_layer = hparams.n_embd_per_layer;++ ggml_tensor * cur = ggml_mul_mat(ctx0, model.per_layer_model_proj, inpL); // [n_embd_per_layer * n_layer, n_tokens]+ cur = ggml_scale(ctx0, cur, 1.0f / sqrtf((float) n_embd));+ cur = ggml_reshape_3d(ctx0, cur, n_embd_per_layer, n_layer, n_tokens);++ cur = build_norm(cur, model.per_layer_proj_norm, nullptr, LLM_NORM_RMS, -1);+ cb(cur, "inp_per_layer", -1);++ // permute to shape: [n_embd_per_layer, n_tokens, n_layer]+ cur = ggml_cont(ctx0, ggml_permute(ctx0, cur, 0, 2, 1, 3));+ return cur;+}diff --git a/src/models/models.h b/src/models/models.hindex 387a4adcb2..1b589ddcf3 100644--- a/src/models/models.h+++ b/src/models/models.h@@ -914,6 +914,23 @@ struct llama_model_gemma_embedding : public llama_model_base {};+struct llama_model_gemma_embedding2 : public llama_model_base {+ llama_model_gemma_embedding2(const struct llama_model_params & params) : llama_model_base(params) {}+ void load_arch_hparams(llama_model_loader & ml) override;+ void load_arch_tensors(llama_model_loader & ml) override;++ struct graph : public llm_graph_context {+ const llama_model & model;++ graph(const llama_model & model, const llm_graph_params & params);++ ggml_tensor * build_inp_per_layer(ggml_tensor * inpL);+ };++ std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;+};++struct llama_model_starcoder2 : public llama_model_base {llama_model_starcoder2(const struct llama_model_params & params) : llama_model_base(params) {}void load_arch_hparams(llama_model_loader & ml) override;diff --git a/tests/test-llama-archs.cpp b/tests/test-llama-archs.cppindex a46e32d7ff..e74d8767cc 100644--- a/tests/test-llama-archs.cpp+++ b/tests/test-llama-archs.cpp@@ -729,7 +729,7 @@ static bool arch_supported(const llm_arch arch) {if (arch == LLM_ARCH_GRANITE_SWITCH) {return false; // FIXME adapter fixture}- if (arch == LLM_ARCH_LLAMA_EMBED || arch == LLM_ARCH_GEMMA_EMBEDDING || arch == LLM_ARCH_T5ENCODER) {+ if (arch == LLM_ARCH_LLAMA_EMBED || arch == LLM_ARCH_GEMMA_EMBEDDING || arch == LLM_ARCH_GEMMA_EMBEDDING2 || arch == LLM_ARCH_T5ENCODER) {return false; // FIXME Embedding (?) models produce inconsistent results.}if (arch == LLM_ARCH_RWKV6 || arch == LLM_ARCH_RWKV6QWEN2 || arch == LLM_ARCH_RWKV7 || arch == LLM_ARCH_ARWKV7) {